Continual Unlearning for Text-to-Image Diffusion Models: A Regularization Perspective

AI summary

Overview

Research area: Machine unlearning for generative models — specifically continual (sequential) concept removal from text-to-image diffusion models, viewed through the lens of regularization and continual learning.

Technical level: Advanced. The paper assumes familiarity with diffusion models, cross-attention key/value projections, gradient/Hessian analysis, and continual-learning regularization literature. The high-level argument is accessible, but the proposed method involves projection matrices and least-squares projections.

One-sentence scope: The paper introduces and benchmarks the problem of unlearning concepts from a diffusion model one request at a time, diagnoses why existing methods collapse, and evaluates a suite of add-on regularizers including a semantic-aware gradient-projection method.

What This Paper Is About

Existing unlearning methods for text-to-image diffusion models assume that all requests to erase concepts arrive at once, but in practice they arrive sequentially — one day a request to remove violent concepts, later a request to remove a copyrighted artist's style. The authors show that applying standard unlearning methods repeatedly in this sequential manner causes "rapid utility collapse": the model still erases targets but forgets everything else and generates degraded images. Their goal is to diagnose the cause of this failure and evaluate practical, plug-and-play fixes that keep the model close to its original capabilities.

Key Contributions

1. First systematic study of continual unlearning (CU) for text-to-image diffusion models. The authors define CU formally with three requirements — erasing new targets, preserving prior unlearning, and retaining unrelated abilities — and build a benchmark extending UnlearnCanvas with style-level and object-level unlearning sequences.

2. Diagnosis of the failure mode. They attribute utility collapse to cumulative parameter drift away from the pre-trained weights, supported by empirical measurements of ℓ2 drift and a Taylor-expansion bound on retention loss.

3. A suite of add-on regularizers compatible with existing unlearning methods: update-norm regularization (L1/L2), Selective Fine-tuning (SelFT) restricted to the top-k% most important parameters, and model merging (TIES-Merging of independently unlearned models).

4. A semantic-aware gradient-projection method. Motivated by the finding that retention difficulty correlates with text-embedding similarity to the unlearned concept, the method projects the unlearning gradient orthogonal to the subspace spanned by semantically close auxiliary concepts, and is shown to be complementary to the other regularizers.

Main Findings

• Popular unlearning methods collapse under sequential requests. ConAbl and SculpMem both perform well on the first request with high UA, RA-I, and RA-C, but as concepts are unlearned sequentially, Unlearning Accuracy stays high while retention accuracy drops drastically, and after unlearning 12 concepts the model fails to generate meaningful content. The paper notes that specific numerical values are shown in figures rather than stated in text.

• Simultaneous unlearning preserves utility but at prohibitive cost. Re-unlearning all prior concepts from scratch at every new request keeps retention higher, but training time grows with the total number of requests.

• Cumulative parameter drift is the root cause. After the first request, sequential and simultaneous strategies show similar ℓ2 drift from the pre-trained weights, but drift grows dramatically with more requests under sequential unlearning while staying nearly constant under simultaneous unlearning. Independently unlearned models also show small, roughly constant shifts.

• A theoretical bound links drift to retention loss. A Taylor expansion around the pre-trained weights bounds the change in retention loss by the gradient norm times the update norm plus half the Hessian norm times the squared update norm; estimated Hessian coefficients were minuscule, indicating a smooth basin.

• Generic regularizers partially fix the problem. L1/L2 update-norm penalties, SelFT, and model merging all reduce drift and improve RA-I and RA-C, with model merging delivering the strongest overall retention in the reported comparison.

• In-domain retention is persistently harder than cross-domain retention. RA-C consistently exceeds RA-I across regularizers, which the authors attribute to concepts in the same domain being semantically closer (e.g., "Bear" vs. "Cat" rather than "Bear" vs. "Van Gogh").

• Retention difficulty correlates with semantic similarity. Unlearning "Abstractionism" produced a strong negative correlation between retention accuracy and text-embedding cosine similarity, and a strong positive correlation between cosine similarity and distortion of key/value vectors.

• Gradient projection gives the best in-domain retention. It achieves the highest RA-I across both style and object unlearning, with slightly lower RA-C than SelFT and model merging in the object case; combining it with other regularizers (measured by the harmonic mean of UA, RA-I, and RA-C) yields further gains.

• Parameter updates are driven by the anchor, not the target. In anchor-based unlearning, distinct targets mapped to the same anchor induce highly correlated parameter updates — suggesting anchor-based unlearning is functionally representation replacement.

• Concept erasure is all-or-nothing. Interpolating between pre-trained and unlearned models leaves outputs visually unchanged over a wide range of coefficients until a critical threshold, after which the concept is abruptly suppressed — complicating model merging.

• Drift is intrinsic to sequential unlearning, not just extra optimization steps. With early stopping at 99% unlearning accuracy and matched cumulative step counts (e.g., 2,100 each after six concepts), sequential unlearning still accumulated substantially more drift than simultaneous unlearning.

• Findings generalize. SculpMem shows the same collapse after 12 concepts in Appendix A, and add-on regularizers improve it consistently; celebrity unlearning with ConAbl is also examined in Appendix A.2.

Methodology in Plain English

The authors start from a pre-trained Stable Diffusion checkpoint provided by UnlearnCanvas, along with specialized classifiers that recognize 60 artistic styles and 20 object categories. This checkpoint can generate all 80 concepts at above 98% top-1 recognition accuracy, giving the authors an objective, standardized way to score whether a concept was successfully erased or successfully retained.

They build two evaluation protocols: unlearn a random sequence of 12 unique artistic styles, or unlearn a random sequence of 12 unique objects. To measure retention they hold out 12 additional styles and 8 objects that are never targeted. Images are generated with the template "A {object} image in {style} style" using 5 random seeds — 40 images per style concept (5 × 8 objects) and 60 images per object concept (5 × 12 styles).

Three metrics are reported: Unlearning Accuracy (UA, how often the classifier no longer returns the erased concept), In-Domain Retention Accuracy (RA-I, for concepts in the same family as erased ones), and Cross-Domain Retention Accuracy (RA-C, for unrelated concepts).

They take two existing unlearning methods, ConAbl and SculpMem, and adapt them in two ways: sequentially, where each new request updates the previously unlearned model, and simultaneously, where the model is retrained from the original weights to unlearn all concepts so far. Comparing the two reveals the drift problem, which they then quantify by measuring the ℓ2 distance between each unlearned model and the pre-trained weights, and by bounding retention loss change via a second-order Taylor expansion.

To fix drift, they study lightweight "add-on" regularizers that plug into the existing unlearning loss: (1) penalizing the L1 or L2 norm of the update relative to the previous checkpoint; (2) SelFT, which ranks parameters by a single-forward-pass first-order importance score and updates only the top-k% (the exact k is not specified in the text); and (3) TIES-Merging of independently unlearned models, which prunes each model's top-k% parameter deviations and averages them, effectively constraining updates to a subspace.

Finally, for semantic awareness they generate auxiliary concepts with an LLM, filter them by text-embedding similarity to the target, and build the span of their embeddings. The unlearning gradient for the cross-attention key and value projection matrices is then projected onto the orthogonal complement of that span, so updates erase the target while making zero first-order change to semantically similar concepts (stated as Lemma 7.1, with proof). The method does not require access to the retain set.

Why This Matters

Impact on research: The paper reframes unlearning as a sequential, drift-accumulation problem rather than a one-shot optimization problem, connecting diffusion unlearning to established continual-learning regularization theory. It also argues that practical progress may come from compatible add-on methods rather than entirely new algorithms, and that the same regularizers help even in the standard single-concept setting.

Real-world applications:

• Regulatory compliance: Regulations such as CCPA grant individuals the right to request removal of personal or copyrighted content, and requests will arrive over time, not in one batch.
• Content moderation pipelines: A service could remove violent or harmful concepts on one day and copyrighted artistic styles on another without degrading the rest of its image generation.
• Personal likeness removal: Erasing a specific person's likeness or a celebrity's identity from a deployed model, as examined in the paper's celebrity erasure experiment.
• Cost-efficient model maintenance: Retraining Stable Diffusion v2 on LAION-5B costs roughly 150,000 GPU-hours per the paper, so sequential unlearning that avoids full retraining has direct practical value.

Industry relevance: Any organization deploying a text-to-image model faces a stream of takedown and safety requests. A method that keeps the model stable across many such requests — rather than one that requires re-unlearning everything from scratch each time — directly reduces compute cost and avoids visible quality degradation for end users.

Future Directions

• Adversarial robustness across sequential steps: Understanding how robustness to adversarial recovery attacks evolves as more unlearning requests accumulate, which the authors identify as critical for safe deployment.
• Generalization beyond the current setting: Whether these challenges compound differently across architectures (e.g., DiT), training objectives (e.g., flow matching), and modalities (e.g., video, speech) beyond diffusion-based image generation.
• Natively sequential unlearning methods: Designing methods that anticipate future requests and account for interactions between them, rather than add-on regularizers applied to one-shot unlearning algorithms.
• Explaining why simultaneous unlearning stays stable: Identifying what property of the simultaneous strategy keeps drift low despite comparable cumulative optimization steps, and using that to design better regularizers.

Target Audience

Researchers and engineers working on machine unlearning, generative model safety, or continual learning — particularly those deploying or maintaining diffusion models where concept-removal requests arrive over time. The paper is also relevant to policy-adjacent readers interested in how "right to be forgotten" requirements translate into technical practice. Readers will need background in diffusion models and optimization to follow the theoretical bound, the projection formulation, and the experimental design; the abstract-level argument about utility collapse is understandable to a broader audience.

Authors’ abstract

Machine unlearning--the ability to remove designated concepts from a pre-trained model--has advanced rapidly, particularly for text-to-image diffusion models. However, existing methods typically assume that unlearning requests arrive all at once, whereas in practice they often arrive sequentially. We present the first systematic study of continual unlearning in text-to-image diffusion models and show that popular unlearning methods suffer from rapid utility collapse: after only a few requests, models forget retained knowledge and generate degraded images. We trace this failure to cumulative parameter drift from the pre-training weights and argue that regularization is crucial to addressing it. To this end, we study a suite of add-on regularizers that (1) mitigate drift and (2) remain compatible with existing unlearning methods. Beyond generic regularizers, we show that semantic awareness is essential for preserving concepts close to the unlearning target, and propose a gradient-projection method that constrains parameter drift orthogonal to their subspace. This substantially improves continual unlearning performance and is complementary to other regularizers for further gains. Taken together, our study establishes continual unlearning as a fundamental challenge in text-to-image generation and provides insights, baselines, and open directions for advancing safe and accountable generative AI.